Beauty is in the Eye of the Beholder: Uncovering Aesthetic Bias in Multimodal Perception and Generation

ACL ARR 2026 January Submission9358 Authors

Keywords: LLM and Society, Bias

Abstract: Beauty standards are not just aesthetic preferences, they are embedded in cultural, social, and economic structures. Yet, as multi-modal AI systems gain widespread influence, from image generation to content curation, their internal aesthetic bias has not been studied before although it possesses great potential to influence the society through their users from all of the world. In this paper, we present a systematic framework to evaluate aesthetic bias in large-scale multi-modal models: Not only do we focus on how models \textbf{perceive} beauty on any given images of human, but also on how generated images from these models reflect certain beauty preference. We introduce a diverse, custom-built portrait dataset alongside a rigorous pairwise comparison protocol to quantify perceptual biases across ethnicity, gender, and aesthetic style. By conducting a large-scale, cross-model evaluation of generated portraits, comparing model outputs to human consensus, we reveal \textbf{consistent and measurable} bias toward certain beauty norms across major models. By surfacing these implicit patterns, our benchmark lays the groundwork for developing more culturally aware, inclusive AI systems, and provides critical insights for researchers, policymakers, and developers aiming to mitigate algorithmic bias in a globally interconnected world.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: bias and fairness, multimodal models, vision-language models, human evaluation, cultural bias, model evaluation

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

Submission Number: 9358